Torus raises $500 million led by Magnetar and unveils modular Foundry inference platform
Summary
Torus, a Salt Lake City-based U.S. company that manufactures distributed power and compute equipment, has secured $500 million in growth capital in a round led by Magnetar. On Nov. 13, 2026, it also introduced Foundry, a prefabricated AI infrastructure product aimed at the rising need for inference compute.
Foundry combines power, storage, cooling, networking and compute in a single factory-built system that is installed at sites that already exist. Torus says each deployment provides between 1–20 megawatts of inference capacity and can move from signed contract to running compute in 12–16 weeks. The system relies on grid power already available at the location rather than new supply. The company contrasts that timeline with the years needed to build a conventional hyperscale campus.
HPCwire carried the announcement. The company describes power access as the main constraint on meeting AI demand and says it has worked on that problem for years. The available text does not name customers, specific sites, the valuation or how the new capital will be allocated.
Why it matters
The pitch targets the slowest part of AI capacity planning: getting new power. By drawing on grid capacity already in place at existing sites, Torus is trying to bypass interconnection queues and multi-year campus construction for inference workloads. A $500 million raise gives the company money to scale manufacturing and deployments. Operators with stranded or underused power at existing facilities may find a new way to monetize it. Customers needing inference capacity quickly could also gain an alternative to waiting on large campus builds.
Expert Lens
The approach hinges on how much spare, deliverable power actually sits at existing sites, and utilities may still require studies or upgrades even when a load stays within existing service limits, which could stretch the stated 12–16 week window. Capacity in the 1–20 megawatt range suits distributed inference better than large training clusters, so the addressable market likely depends on how latency-sensitive inference demand evolves. Packing power, cooling and compute into prefabricated units also raises questions about thermal density, serviceability and refresh cycles as GPU generations turn over. Lenders and offtakers may treat these modular assets differently from long-lived campus real estate, which could shape financing terms. The next signal to watch is whether Torus announces named customers and specific site deployments that confirm its contract-to-compute timeline in practice.
Sources
This brief was drafted with AI assistance from public sources; every number is attributed to the sources above. See the AI-assisted Disclosure and Methodology.